Verify the final stream cleanup head and close the remaining evidence draft

Written 2026-10-21 15:38 UTC. RE-VERIFY: query each linked PR head/state and runner result before acting. This is a snapshot. The assigned fix round has a two-hour time box ending around 15:40 UTC; code and focused reviews are complete, final full-runner verification is pending. Local watchers were stopped and verified absent; the clean checkout, generated probe source/jars, and any matching checkout cache were removed at 15:36 UTC. GitHub jobs continue independently.

Design decisions in force

Assigned 2026-10-21: "R1. Close finishes within the existing bounded waits even if a write on the shared connection is stalled"; "R2. No thread that is writing part of a frame on the shared connection is ever interrupted by stream cleanup — upload pushes AND output pulls/any other writer"; "R3. Everything the PR already guarantees still holds (echo of an in-flight upload completes or fails clearly; re-entrant close releases exactly once; source-read waits still woken)." No new or raised timeout, no merges/enqueue, no publication/deployment, no weakened/disabled/deleted tests, no reflection/mocks. Finish at a green bld-build, not merge. No subagents are permitted by the common brief.

Mission summary

The original user requested extreme-condition end-to-end tests that fail before bugs are fixed. This round addresses review findings in https://github.com/CodexCoder21Organization/UrlResolver/pull/1371, originally at 6235af9bfc0b04a88d12237ea144dd73e30496cc. Earlier completed work is recorded at https://www.handoff.wasmserver.com/handoffs/hf-2026-10-21-verify-and-fix-closing-an-in-flight-echo-upload-without-losing-the-shared-connection . This checkpoint covers the later review findings and the additional related failure exposed by this round's full runner.

What was found and done

1. Close sent unfinished upload releases synchronously before bounded joins. A full custom controller output held the shared write lane, so release could never acquire it and close never reached a bound. The fix gives release an uninterrupted daemon sender, waits within the existing five-second worker allowance, reports expired/unstarted releases, and reports later sender failures. A genuine 1024-byte bounded-pipe regression holds the upload frame unfinished through cleanup. A balanced public idle hold controls unrelated idle retirement; request timeouts stay at 1000 ms.
2. Output pull workers were interrupted after their joins even while writing a partial __stream_pull frame. Those interrupts are removed. The TCP regression pauses after byte one, holds that frame through close, checks zero interruptions/original controller retention, then delivers all bytes and a health RPC using the same connection.
3. The first full fixed runner passed 2298/2298 final results but recorded a failed first attempt in unchanged testRpcStreamServerHandlerErrorDoesNotDeadlock. This was NOT called unrelated: both current and preceding main artifacts were clean. The original evidence artifact also failed that test all three attempts. A delayed push can reach the service after its handler failed and removed the target; the sender waits for registration, and close unnecessarily spends its upload join on that RPC sender after release returned. A new deterministic public-API test forces exactly this ordering and failed 3/3. The final change joins only workers still stopping source reads, under their readLock; active RPC senders finish/fail through their sender without cleanup waiting or interruption.
4. Final source was compiled against published real resolver dependencies and passed 22/22 focused checks: late-push 7/7, original R1/R2 each 3/3, all six original/neighbor cases, and the unchanged handler-error test 3/3. Separate code and test review passes are complete. All original tests remain byte-identical. No timeout/iteration/assertion changes.

Relevant PRs / refs

No source is merged/enqueued by this round. No production deploy, server restart or Maven publication was performed. No local-only code, uncommitted changes or stashes remain.

Evidence

• Original R1/R2 failure runner: https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/39613269155/job/119267478762 . Artifact https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/39613269155/artifacts/12356172635 has 2298 tests, 2295 final passes and 3 failures; each requested regression and the unchanged handler-error case fails all 3 attempts with the intended assertion.
• Intermediate head beca9d63b730484c79f55c8b41d222f07922e5dd required check passed at https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/39613911072/job/119269704845 . Artifact https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/39613911072/artifacts/12355521850 has 2298 final passes; all five original/new stream tests passed first attempt, but handler-error failed once then passed. This is historical validation, NOT the final head's required result.
• Earlier R3 fail-first artifact https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/39570691835/artifacts/12334699955 was independently checked: original upload-frame and reentrant-close tests each failed 3/3; source-read cancellation already passed and is a preserved behavior.
• Main artifacts checked: https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/39598742185/artifacts/12348329319 (2293 first-attempt passes) and https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/39592579931/artifacts/12346435262 (2290 first-attempt passes).

Next steps

1. Re-verify current state first, especially exact final head and effective branch rules. Claim this handoff and clear the time-box blocked reason when resuming. Only bld-build is required; both kotlin.build App checks are informational under the current common brief.
2. Watch final https://github.com/CodexCoder21Organization/UrlResolver/pull/1371 with plain build-watchman 0.0.21. Do not merge/enqueue. Inspect the final attempt artifact: the new late-push case and unchanged handler-error case must be first-attempt clean, as must the original stream regressions. Record full suite counts and any failed-attempt history honestly.
3. Collect the late-push draft runner at https://github.com/CodexCoder21Organization/UrlResolver/pull/1379 . Expected failure is testClosingReleasedUploadDoesNotWaitForLatePush at the cleanup wait assertion, not setup/compile. Record it and close the draft with no merge.
4. Update the real PR validation body with the final passing runner link and the second evidence result. Complete/report/archive this handoff when the assigned fix round is fully verified. If still running at time box, release the claim and leave this checkpoint current.
5. Previous local watchers and checkout/probe outputs are already removed; no matching checkout cache remained. Any continuation should create a fresh checkout if needed and own its own plain watcher. No further code change or extra verification batch is planned unless the final runner exposes a real problem.

Reusable / operational knowledge

• Common instructions explicitly require GitHub-runner verification because this box cannot reliably run kompile. Required workflows build and test with --local-and-remote. Supplementary probes compile exact manager/report sources and test bodies against published resolver 0.0.1328 plus real dependencies; they are not a rebuilt full module. The final full runner remains mandatory.
• Old gh pr edit fails on deprecated projectCards GraphQL. Write a JSON body file and PATCH repos/CodexCoder21Organization/UrlResolver/pulls/1371 through gh api. Closing the first draft used PATCH state=closed. No challenge CLI was invoked because it automatically enqueues/merges, forbidden here.
• gh run view --log-failed sometimes returns empty output for completed jobs. Raw job-log API and attempt artifacts work. Download with gh api repos/CodexCoder21Organization/UrlResolver/actions/artifacts/<id>/zip. XML root has test nodes with name/success/time and attempt-history nodes containing attempt nodes; inspect failed attempts, not just final aggregate.
• Watcher GitHub connect timeouts recovered on later reads; direct gh CLI also succeeded. No required job was rerun. Superseded tests-only draft revisions were cancelled before the final evidence head ran; neither a required failed job nor a merge queue was rerun.
• Working directory: /tmp/claude-1000/-code/15741e5f-c64b-403b-8e66-397a9b1d44d9/scratchpad/workspace/s42-root/workspace/UrlResolver . Findings: /tmp/claude-1000/-code/15741e5f-c64b-403b-8e66-397a9b1d44d9/scratchpad/cx/out/s42-findings.md . Build jars, caches and classpath files are transient outputs and intentionally excluded from source branches.

Durable supporting evidence and the XML reader are preserved at https://github.com/CodexCoder21Organization/UrlResolver/tree/wip/s42-review-evidence-20261021/handoffs/artifacts/s42 . Full raw XML remains in the linked Actions artifacts; transient compiler jars/classes/caches are intentionally omitted. Final runner job: https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/39624401576/job/119305900593 .

Final brief audit: R1 and R2 are implemented with original runner fail-first evidence and intermediate full-run passing evidence. R3 original scenarios are preserved, and the related late-push case is fixed with 3/3 focused baseline failures and 7/7 focused fixed passes, plus the unchanged handler-error test passing 3/3. Final head has 22/22 focused passes and completed separate reviews. Required bld-build on the final head and the late-push draft runner are STILL RUNNING, so the assigned fix round is not yet fully verified. No existing tests were edited; no merge/enqueue/publication/deployment occurred.